TL;DR: OpenRouter 's token markup is genuinely 0%, but loading credits by card costs 5.5% with a $0.80 minimum, crypto costs about 5%, and BYOK costs 5% past 1M requests a month (vendor FAQ, verified July 2026). If the deposit fee is your problem, Vercel AI Gateway is the closest 0%-markup swap. If one provider covers 90% of your spend, go direct. If you want $0 fees at scale and own your infra, self-host LiteLLM. The routing table below matches the fix to the pain.
Most " OpenRouter alternatives" lists are written by the alternatives. Three of the pages ranking for this query in July 2026 are vendor blogs that rank themselves first. This page takes the buyer's side: every fee below was checked against a live vendor pricing page or docs page in July 2026, every user complaint is labeled as a user report rather than our measurement, and each alternative comes with the catch its own marketing omits.
Start with why you are leaving, because OpenRouter 's actual pricing is better than its reputation and worse than its homepage. Our OpenRouter review confirmed the 0% inference markup — you pay the same per-token price the provider lists. The costs sit elsewhere, in the vendor's own FAQ (verified July 2026): a 5.5% fee with a $0.80 minimum on Stripe credit purchases, roughly 5% on crypto, and bring-your-own-key routing that is free for the first 1M requests a month, then 5% of what the same traffic would cost on OpenRouter credits. In dollar terms: a $1,000 top-up loses $55 before the first token, and the $0.80 minimum turns a $5 top-up into a 16% effective surcharge — the fee math that dominates third-party complaint threads (user reports, not our measurements). Add the structural issue no fee schedule shows: one gateway between you and every model is one shared point of failure. StatusGator, a third-party monitor, records 46+ OpenRouter outage events over the past year, with the last acknowledged incident on February 19, 2026 (StatusGator's count, verified July 2026; the monitor does not attribute root causes, so we cannot say how many events were gateway-wide versus isolated upstream-provider issues).
Those are the three exits, and each leads somewhere different: the deposit fee points to a 0%-fee gateway, the dependency risk points to your own proxy or a direct API, and the upstream-quality lottery points to a first-party open-model host. Route yourself accordingly.
Disclosure: we have no affiliate or business tie to OpenRouter , Vercel, Anthropic, LiteLLM, Portkey, Requesty, Together AI, or Fireworks AI as of publication; if that changes, this line will say so. We have not run our hands-on gateway suite on Requesty, Portkey, Fireworks, or LiteLLM yet — where a claim below is a vendor number or a user report rather than our test, it is labeled as such.
Every alternative at a glance
- Vercel AI Gateway — 0% markup, per-request ZDR on paid plans
- Direct provider APIs — when 1–2 providers are enough (Anthropic batch 50%, cache 0.1×)
- LiteLLM — your keys, your logs, your uptime
- Portkey / Requesty — middle ground with observability
| # | Alternative | Fee structure (verified July 2026) | Best for | The catch |
|---|---|---|---|---|
| 1 | Vercel AI Gateway | 0% token markup on paid tier, free BYOK; prepaid credits carry processing fees; governance add-ons from $0.10/1k req (vendor docs) | The deposit-fee escape with the same catalog shape | Vercel account lock-in; failed BYOK calls silently retry on system credits and bill you |
| 2 | Direct provider APIs | $0 platform fee; provider list price only, e.g. Claude Sonnet 5 $2/$10 per MTok intro until Aug 31, 2026 (vendor price) | Teams where one model family is 90%+ of spend | One key = one vendor; no cross-provider fallback |
| 3 | LiteLLM (self-hosted) | Open-source proxy free; enterprise tier custom (vendor) — total cost is provider prices + hosting + ops time | $0 fees at any scale, compliance-bound infra | You become the gateway team: deploy, patch, monitor |
| 4 | Portkey | Free to 10k logged req/mo; $49/mo for 100k + $9 per extra 100k (vendor price) | Production governance: logs, retries, guardrails | A subscription regardless of usage; doesn't fix the fee pain |
| 5 | Requesty | 5% markup on model costs, pay-as-you-go, no minimums (vendor price) | EU data residency in an OpenRouter-shaped drop-in | 5% on tokens ≈ 5.5% on deposits at steady spend |
| 6 | Together AI | No stated platform fee; open models at list, e.g. Llama 3.3 70B $1.04/$1.04 per MTok (vendor price) | First-party control over open-model serving | Open models only — no Claude, GPT, or Gemini |
| 7 | Fireworks AI | No platform fee; DeepSeek V4 Flash $0.14/$0.28 per MTok; cache 50% off, batch 50% off (vendor price) | Cheapest stacked discounts on open models, postpaid | Same closed-model gap; pricing scattered across pages |
Fees last verified July 2026 against openrouter.ai/docs/faq, vercel.com/docs/ai-gateway/pricing, platform.claude.com pricing docs, litellm.ai, portkey.ai/pricing, requesty.ai/pricing, together.ai/pricing, and fireworks.ai pricing pages plus its serverless pricing docs, all fetched directly. Per-model token prices move often — our live LLM price tracker is the current source.
1. Vercel AI Gateway: the deposit-fee escape
The most direct answer to " OpenRouter without the 5.5%." Vercel's AI Gateway pricing docs (page dated June 20, 2026, verified July 2026) state 0% markup on tokens on the paid tier, including BYOK — "no markup or fee from AI Gateway" — with the same multi-provider catalog shape you left: one API key, many models, automatic fallback. Zero-data-retention is available per request at no cost on Pro and Enterprise, which is a cleaner privacy posture than OpenRouter 's opt-in logging discount (more on that below).
Read the fee print before declaring victory. You pay via prepaid credits and, in Vercel's words, "you're responsible for any payment processing fees" — so "no fee" is not quite $0, it is "no Gateway fee." Governance features are metered add-ons: team-wide provider allowlists or team-wide ZDR enforcement run $0.10 per 1,000 requests, and custom reporting costs $0.075 per 1,000 writes plus $5 per 1,000 queries. The free tier covers only a subset of models at low rate limits. Two structural catches: the whole thing ties you to a Vercel team account and its ecosystem, and — the one that surprises people — when a BYOK request fails, the Gateway silently retries on system credentials and bills your credit balance. If you came here to escape surprise charges, put an alert on that balance.
2. Direct provider APIs: when one vendor is enough
If 90%+ of your spend already lands on one model family, the gateway is pure overhead. Going direct to Anthropic, OpenAI, or Google means $0 platform fees, no deposit dance, no middle layer to go down, first-party prompt caching and batching at full fidelity, and real invoicing at enterprise tiers — the thing user reports most often flag as missing from OpenRouter 's credit model.
The anchor numbers (vendor price, platform.claude.com pricing docs, verified July 2026): Claude Sonnet 5 costs $2/$10 per million tokens in/out at its introductory rate through August 31, 2026, rising to $3/$15 after; Opus 4.8 is $5/$25; Haiku 4.5 is $1/$5. The Batch API halves those, and cache reads cost 0.1x input price. Comparable first-party discounts exist at OpenAI and Google — the full cross-provider table lives in our live price tracker .
That Sonnet 5 date is also the honest counterargument. On September 1, 2026, its price rises roughly 50%, and a single-provider shop just eats it — no router to shift traffic elsewhere. Going direct also means separate billing, keys, and SDKs per provider if you ever use several, plus per-provider rate-limit tiers you must grow into. Direct APIs are the right call for concentration, not for diversification.
3. LiteLLM: $0 fees at scale, if you run it yourself
LiteLLM is the only option on this list where platform fees stay $0 at $10,000 a month of spend, because there is no platform — it is an open-source proxy you host, pointing at your own provider keys at list prices. The project claims 240M+ Docker pulls and 1B+ requests served (vendor claims, litellm.ai, verified July 2026, not independently audited), with 100+ provider integrations, per-team budgets, guardrails, load balancing, and air-gapped deployment for compliance-bound teams. An enterprise tier exists at "Get In Touch" pricing with a 30-day trial.
The total cost is not $0; it is provider prices plus hosting plus the line item vendors never print — your ops time. You become the gateway team: deploy it, upgrade it, monitor it, patch it, and debug it at 2am when routing breaks. The recurring theme in third-party comparison posts and community threads (user reports, unverified individually) is upgrade churn and configuration complexity. We have not run our hands-on suite on LiteLLM yet. Whether owning this layer beats renting it is exactly the calculus in our build vs buy guide for AI agents : below a few hundred dollars a month of LLM spend, OpenRouter 's fee is cheaper than one engineer-hour; well above it, LiteLLM's math takes over.
4. Portkey: governance, not fee relief
Portkey solves a different problem than the one that probably brought you here. Vendor pricing (portkey.ai/pricing, verified July 2026): a free Developer tier with 10,000 logged requests a month, 3-day log retention, and no overages; Production at $49 a month for 100,000 logged requests plus $9 per additional 100,000 up to 3M; an open-source gateway you can self-host free; and custom Enterprise with SOC 2 Type 2, GDPR, HIPAA, and VPC deployment.
Note the pricing unit: per logged request, not a token markup — so heavy traffic with light logging stays cheap, and the fee scales with observability, not with model spend. That is the right shape for teams that outgrew "just routing" and now need production logs, retries, fallbacks, guardrails, and dashboards. The catch runs the other way: it is a subscription you pay regardless of usage, capped by logged-request tiers, and it does nothing about deposit-fee pain because it is not trying to. It is also a heavier product to adopt than a drop-in router. We have not run our hands-on suite on Portkey yet.
5. Requesty: the EU-residency drop-in
Requesty is the most OpenRouter -shaped thing here: 400+ models behind one API, pay-as-you-go, with EU data residency — the feature that gets European teams past their own procurement. Vendor pricing (requesty.ai/pricing, verified July 2026): a flat 5% markup on base model costs, no subscription, no seat fees, no minimum spend; a free tier of 200 requests a day on free models; Enterprise on request with SSO, RBAC, and PII detection.
Do the arithmetic before switching for the fees alone: 5% on every token is roughly a wash with OpenRouter 's 5.5% deposit fee at steady spend. What Requesty actually fixes is the fee structure — one clean 5% line on usage instead of deposit minimums that punish small top-ups — not the fee's existence. And it is a smaller company with a thinner public track record than anything else on this list; we have not run our hands-on suite on it yet.
6. Together AI: first-party open-model serving
One of OpenRouter 's most persistent user-report complaints is upstream variance: the same model name can route to different providers running different quantizations, so quality shifts run to run (a theme across community threads and third-party comparisons; we have not reproduced it in our own testing). Together AI removes the lottery for open-weight models by being the provider: you get first-party control of serving speed and quantization, plus a growth path to dedicated GPUs (H100s from $5.49/hr).
Vendor prices per million tokens in/out (together.ai/pricing, verified July 2026): Llama 3.3 70B at $1.04/$1.04, DeepSeek V4 Pro at $1.74/$3.48 with $0.20 cached input, Kimi K2.7 Code at $0.95/$4.00, gpt-oss-20B at $0.05/$0.20. No stated platform fee. Kimi K2.7 Code is a serious budget coding pick — context in our best LLM for coding roundup, and the cross-provider budget math is in cheapest LLM API .
The catch defines the category: open models only. No Claude , no GPT, no Gemini — so Together replaces the open-weight slice of your OpenRouter traffic, not all of it. And you are back to a single provider, the exact dependency shape you may have adopted OpenRouter to avoid.
7. Fireworks AI: the discount stack
Fireworks competes on latency and on the best stacked discounts on this list. Vendor prices per million tokens in/out (fireworks.ai/pricing plus docs.fireworks.ai/serverless/pricing, verified July 2026): DeepSeek V4 Flash at $0.14/$0.28, DeepSeek V4 Pro at $1.74/$3.48, Kimi K2.7 Code at $0.95/$4.00, gpt-oss-120B at $0.15/$0.60, with unlisted models over 16B parameters at $0.90 per MTok. Cached input is 50% off and batch is another 50% off — stack both on a cacheable batch workload and the effective rate drops far below anything a gateway resells. Billing is postpaid with high rate limits: no prepaid credit dance at all, which for some teams is the entire reason to leave OpenRouter . Dedicated H100s run $7/hr on-demand.
The catches: the same closed-model gap as Together; the $1 of free credit is a token gesture, not an evaluation budget; and the pricing is scattered between the marketing page and the docs, which is why we cite both. We have not run our hands-on suite on Fireworks yet.
Pick by situation
| Your situation | Switch to | Why |
|---|---|---|
| The 5.5% deposit fee is the problem; you want the same multi-model catalog | Vercel AI Gateway | 0% markup and free BYOK on the paid tier; per-request ZDR free. Watch the BYOK silent-retry billing. |
| One provider is 90%+ of your spend | Direct API (Anthropic / OpenAI / Google) | $0 platform fee, first-party caching and batching, real invoices. Accept single-vendor pricing risk — see Sonnet 5's Sept 1 increase. |
| $0 fees at real scale, and you have engineers | Self-hosted LiteLLM | Provider list prices on your own keys forever; you pay in ops time instead. |
| You need logs, guardrails, and audit trails more than routing | Portkey | Priced per logged request; SOC 2 Type 2 / HIPAA at Enterprise. Doesn't fix fee pain. |
| EU data residency, minimal migration | Requesty | OpenRouter-shaped drop-in with one flat 5% line. You trade fee structure, not fees. |
| Open models only, tired of the upstream-provider lottery | Together AI or Fireworks AI | First-party serving control; Fireworks stacks 50% cache + 50% batch discounts, postpaid. |
| You spend under ~$100/month | Stay on OpenRouter, probably | The fee is dollars a month; see the section below before migrating anything. |
One privacy footnote that cuts across the table. OpenRouter does not train on your prompts and does not log them by default (vendor policy, verified July 2026), but it offers a 1% discount for opting into prompt logging — and its terms of service grant OpenRouter broad commercial-use rights over logged inputs and outputs, which drew visible pushback on Hacker News. OpenRouter has since added a zero-data-retention routing setting. If ZDR is a requirement rather than a preference, note that Vercel AI Gateway includes per-request ZDR at no cost on paid plans, with team-wide enforcement as a metered add-on. We found no litigation on any of this as of July 2026.
Where leaving OpenRouter falls short
The honest section: for a lot of readers of this page, the right move is staying put.
At small spend, the fee is noise. If you top up $100 a month, the 5.5% Stripe fee costs $5.50 — less than the coffee consumed during the migration meeting. Rewriting routing code, revalidating model behavior, and re-learning a new dashboard costs real engineering days. The fee math only justifies a move when the monthly fee exceeds the amortized migration cost, which for most solo developers and small teams it does not. (The one cheap fix inside OpenRouter : stop making small top-ups. The $0.80 minimum is what turns $5 deposits into a 16% surcharge; larger, less frequent deposits pin the fee at 5.5%.)
The core product is still competitive. 0% token markup is real, the catalog is the broadest in the category, and pass-through pricing means our price tracker numbers apply unchanged. Several alternatives above charge more in practice: Requesty's 5% on tokens, Portkey's subscription, LiteLLM's ops payroll.
Most alternatives reintroduce the dependency, one layer over. Leave OpenRouter for Vercel AI Gateway and you depend on Vercel. Go direct to Anthropic and you depend on Anthropic — with a scheduled 50% price increase on Sonnet 5 as the live demonstration. Together and Fireworks are single providers by definition. The only option that removes the dependency layer rather than relocating it is self-hosted LiteLLM, and it charges you in operations instead. On the outage record itself, keep StatusGator's 46+ events in proportion: the monitor counts events without attributing root causes, and a multi-provider gateway inherits every upstream provider's incidents — the same providers you would depend on directly after leaving.
And the caveat on us: we have not run our hands-on gateway suite on Requesty, Portkey, Fireworks, or LiteLLM. Everything above about them is verified vendor pricing plus labeled user reports. Hands-on latency and failover tests are budgeted; this page will get a changelog entry when they land.
All guides in this topic
- OpenRouter review — the fee-by-fee breakdown this page spins off
- Live LLM API price tracker — current per-model prices across providers
- Cheapest LLM API — the budget path, including Together and Fireworks
- Best LLM for coding — where Kimi K2.7 Code and the frontier coding models land
- Build vs buy for AI agents — the self-host decision behind the LiteLLM option
- Claude vs ChatGPT — if the real question is which single provider to go direct with
No comments yet